YerevaNN is looking for people who want to devote their career to scientific research in computer
science.
YerevaNN is a research lab focused on machine learning. It was founded in 2016 and started to collaborate
with researchers from the University of Southern California. This collaboration resulted in a paper on
electronic health records in Nature
Scientific Data and another workshop paper on biological
NLP in 2019. In 2019 we set a goal to have at least three strong research paper submissions
every year and reached this goal in 2021. Currently we are looking to enlarge the team and target twice
more research output in two years. You can learn more about our research on our website.
We have five projects where new researchers can be immediately included:
- Machine learning for cheminformatics. How well can deep learning algorithms predict chemical
properties of a molecule given its structure? This
paper gives a nice overview. (urgent, collaboration with Institute of Molecular Biology, NAS
RA)
- Cross-lingual NLP. Can a text classifier trained on English data work on other languages
without more training data? A novel algorithm developed in YerevaNN called WARP allows few-shot learning in one
language. Can it perform few-shot or even zero-shot cross-lingual transfer? (collaboration with USC
ISI, US)
- Domain generalization. How can a neural network trained on data from five hospitals work well
on data from the sixth, previously unseen
hospital? We developed diagnostics tools that highlight failure
modes of existing domain generalization algorithms. Can these tools help develop better algorithms?
(collaboration with USC ISI, US)
- Robust estimation. We have determined the
necessary conditions when the matching map between two sets of feature vectors can be correctly
recovered when one of the sets contains outliers. Is it possible to extend to the case when there
are outliers in both sets? Is it possible to recover the true matching map given linearly
transformed data? What are the possible applications of this problem? (urgent, collaboration with
ENSAE/CREST, France)
- Optimization for distributed machine learning. How can many distributed machines, each owning
its own local training data, collaboratively train a single machine learning model without revealing
their local data and communicating as little information as possible to each other? How fast and how
efficient can a distributed optimization method achieve consensus across the machines? This paper is a recent related
work. (with King Abdullah University of Science and Technology, Saudi Arabia)
There are more projects which are not started yet and require a more experienced person to initiate the
collaboration:
- Machine learning for healthcare (with St. George Hospital, UK)
- Reinforcement learning for agriculture (with University of Illinois Urbana Champaign, USA)
- Combinatorial optimization with ML elements (with Khalif University, UAE)
- Reinforcement learning for computer games (with University College London, UK)
- Machine learning for Electronic Design Automation
Senior researchers can propose and lead their own projects in ML. We will help in finding students
and forming a small team.
We expect full time commitment to the work except for the time required for studies (if you are a
student) and the time required for teaching (if you teach). We strongly encourage our team members to
teach. We have great relationships with basically all training centers and universities in Armenia that
have machine learning courses and occasionally get offers to teach there.
Benefits:
- Research environment with constant flow of ideas.
- Salary, depending on the skills and experience, will be between 200K and 1.5M.
- People who have passed a couple of introductory courses in machine learning and have
completed one or two projects can expect a 200K initial salary.
- The salary will be higher for those who have experience in performing literature review,
designing experiments, and writing papers.
- Those who can independently lead research projects, supervise students, write and proofread
papers, initiate collaborations with other teams, and write grants, can expect a 1.5M+
salary.
- One of the strongest health insurance packages in Armenia.
- YerevaNN pays for participation in conferences and summer schools.
- Flexible working hours.
Requirements:
- Strong desire to work on unsolved scientific problems, read and write papers.
- Good knowledge of Python.
- Good knowledge of PyTorch framework. If your previous experience is in TensorFlow, we will ask you
to
switch to PyTorch, as our entire codebase is in PyTorch.
- Good knowledge of English as we collaborate with researchers from many countries.
Please apply using this
form by September 1, 2021.